add comparison

compare.py

@@ -0,0 +1,6 @@
+import torch
+
+logits_cpu = torch.load("logits_cpu.pt")
+logits_vitis = torch.load("logits_vitis.pt")
+
+print("Mean relative diff", ((logits_cpu - logits_vitis).abs() / (logits_cpu.abs() + 1e-12)).mean())

logits_cpu.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8134f12382c238ce288a439bc4aa19359b76393fc5e8ab9b092d0b4848d2551
+size 1006635

logits_vitis.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:453e07ecbc0c63659f5ecf6dc3a84dc9943d3fce81d6383149e406bebf235d6a
+size 1006645

run_model.py

@@ -1,13 +1,18 @@
 from optimum.onnxruntime import ORTModelForCausalLM
 from transformers import AutoTokenizer
+import torch
 
-# provider_options = {"config_file": ""}
+provider_options = {"config_file": r"C:\Users\Felix\HF\RyzenAI-SW\example\opt-1.3b\opt-onnx\vaip_config.json"}
+#provider_options = None
 
-model = ORTModelForCausalLM.from_pretrained(".", provider="CPUExecutionProvider")
+#provider = "CPUExecutionProvider"
+provider = "VitisAIExecutionProvider"
+
+model = ORTModelForCausalLM.from_pretrained(".", provider=provider, provider_options=provider_options)
 tokenizer = AutoTokenizer.from_pretrained(".")
 
 inp = tokenizer("Once upon a time", return_tensors="pt")
 
-res = model.generate(**inp, max_new_tokens=15)
+res = model(**inp)
 
-print(tokenizer.batch_decode(res))
+torch.save(res.logits, "logits_vitis.pt")